Source: Ars Technica
OpenAI is implementing output-level guardrails against exact stylistic replication, offering "feeling" approximations instead of precise author voice reproduction. The move is a liability hedge: the company distinguishes between learning from training data (which it argues is protected) and on-demand reproduction (which courts may view as infringement). It exposes how fragile the "it's just statistical patterns" defense is when users can directly request what amounts to unauthorized derivative works.